Hy3 is Tencent Hunyuan's production-grade Mixture-of-Experts model — 295B total parameters with only 21B active per pass (192 experts, top-8 routing), the upgraded release built on the Hy3-preview line. It expands RL-training scale and post-training data quality for further gains in reasoning, long-context, and agentic tasks, reaching results comparable to flagship models several times its parameter size. It serves a 256K-token context window (text in, text out) with configurable reasoning effort, and is built for real-world coding, tool use, and multi-step agent workflows at a strong quality-to-cost ratio.
Tencent Hy3 is a text-only large language model developed by Tencent. It is designed to process and generate text with a context window of 262,144 tokens, enabling it to work with very long documents…
Tencent Hy3 excels at any task that involves processing large amounts of text in a single context. This includes summarizing entire books, extracting information from long reports, analyzing legal contracts end-to-end, and reviewing large software codebases for bugs or improvements. It is also strong at maintaining coherence in long-form generation tasks such as writing detailed articles or multi-turn dialogues with extensive history. Its BrowseComp score suggests it is particularly good at tasks that require navigating and understanding large amounts of textual information, similar to browsing a document collection.
You should choose Tencent Hy3 when your task requires a large context window that smaller models cannot accommodate without chunking. For example, if you need to analyze a 200-page document in one go or maintain conversation context over hundreds of turns. For simpler tasks like single-paragraph classification or translation, a smaller, cheaper model would be more cost-effective. Tencent Hy3 is also a good choice when you need predictable, transparent pricing without overhead. If your task is multimodal, you will need a different model that supports images or audio.
The model’s primary strength is its ability to handle up to 262,144 tokens in a single prompt, which is one of the largest context windows available. This eliminates the need for complex chunking strategies for many long documents. It scored 84.2 on BrowseComp, indicating robust performance on tasks that involve browsing and extracting information from long textual sequences. The model is likely to maintain accuracy and coherence across very long inputs, though exact performance benchmarks for other tasks are not provided. Its strengths are best realized in scenarios where the entire document must be considered at once.
BrowseComp is a benchmark designed to evaluate a model's ability to browse and comprehend large amounts of text, simulating tasks like searching for specific information across multiple documents or long sections. A score of 84.2 indicates strong performance in these types of tasks. While the exact metric is not detailed, this suggests that Tencent Hy3 is effective at retrieving and synthesizing information from long contexts. For comparison, a score above 80 is generally considered very good. This benchmark is particularly relevant for applications like research assistants and document analysis.
Specific latency figures for Tencent Hy3 are not provided. As a large model with a 262k context window, inference time will depend on input and output token lengths, hardware, and current load. In general, longer inputs increase processing time. Users should expect higher latency compared to smaller models for very long documents. OrcaRouter's API may offer different endpoints with varying speeds. For real-time applications, it is advisable to test with representative inputs. The model’s speed-accuracy trade-off is typical for its size.
The only provided benchmark is BrowseComp, where it scored 84.2, indicating strong long-context browsing and comprehension. Strengths likely include accurate information retrieval from long texts and maintaining context over many tokens. Weaknesses are not explicitly given, but as a text-only model, it lacks multimodal capabilities. Other benchmark scores are not provided, so we cannot compare on standard LLM tasks like reasoning or math. Its large context may introduce higher latency and computational cost compared to smaller models. The model is designed for text-intensive tasks rather than general-purpose chat.
With a context window of 262,144 tokens, Tencent Hy3 is built to process long sequences in a single pass. This means that information at the beginning of a long document can still be available when generating output at the end, reducing retrieval errors common with chunking. The BrowseComp score of 84.2 suggests the model performs well on tasks requiring attention to details spread across long text. Handling very long sequences may still require careful prompt design to avoid overloading the model, but the large window provides significant headroom for most practical long-document use cases.
Through OrcaRouter, billing is transparent and based solely on token usage. You pay $0.18 per one million tokens sent to the model (input) and $0.59 per one million tokens generated by the model (output). These rates are exactly what the provider charges, with no markup added by OrcaRouter. Token counts are computed by Tencent’s tokenizer. There are no subscription fees or minimum usage requirements. Charges appear on your OrcaRouter invoice. This simple structure allows you to predict costs for tasks of known length.
For long-context tasks, Tencent Hy3’s pricing is competitive for the context window size. Many models with similar context windows have higher per-token costs. However, if your task only requires a small context, a cheaper model would be more economical. The zero-markup policy means you are not paying extra for the platform. The cost per million tokens is fixed, so processing a 262k token input costs roughly $0.047 for input (since 262k is 0.262M, input cost 0.262 * 0.18 = $0.047). Output tokens add proportionally. This makes it affordable for many applications.
No specific discounts or caching mechanisms are mentioned for Tencent Hy3 through OrcaRouter. The pricing is per-request based on token count, with no volume discounts announced. Caching of prompts or responses is not described; each request is likely processed independently. Users should assume no special pricing optimizations beyond the base rate. For high-volume usage, it may be worth contacting OrcaRouter for potential custom arrangements, but none are guaranteed. The advertised $0.18/$0.59 per million tokens is the standard public rate.
To call Tencent Hy3, set the base URL to https://api.orcarouter.ai/v1 and use the model identifier "tencent/hy3". The API is fully OpenAI-compatible, so you can use the OpenAI Python client or any HTTP client that supports the Chat Completions endpoint. For example, with the OpenAI Python library, set openai.api_base = "https://api.orcarouter.ai/v1" and include your API key. Then create a chat completion using model="tencent/hy3". The request format is identical to OpenAI's, with messages array, temperature, max_tokens, etc. No special parameters are required beyond the model name.
The API supports all standard parameters of the OpenAI Chat Completions endpoint. These include 'messages' (required), 'temperature' (0–2, default 0.7), 'max_tokens' (limit on output length), 'top_p', 'frequency_penalty', 'presence_penalty', 'stop', and 'stream' for token-by-token output. You can also set 'seed' for deterministic sampling if desired. The API will respect the 262,144-token context limit; prompts exceeding that will be truncated or rejected. There is no special 'multimodal' parameter because the model is text-only. The response format follows OpenAI's structure.
Yes, migrating from another text LLM to Tencent Hy3 via OrcaRouter is straightforward due to the OpenAI-compatible API. You only need to change the model identifier from your previous model (e.g., "some-other-model") to "tencent/hy3" and update the base URL to https://api.orcarouter.ai/v1. The request and response schemas are identical. You may need to adjust prompt formatting if the previous model had different behavior, but typically no code changes are required beyond the model name. It's advisable to test with representative inputs to confirm output quality.
Authentication uses an API key provided by OrcaRouter. You must include this key in the Authorization header as "Bearer YOUR_API_KEY" when making requests. The API key should be kept secret and not shared. OrcaRouter may offer different tiers of keys with varying rate limits. For the OpenAI Python client, set openai.api_key = "your-key". No other authentication is required. Ensure you are using the correct base URL (https://api.orcarouter.ai/v1) and that the key has permissions to access the "tencent/hy3" model.
Compared to models with smaller context windows (e.g., 4k, 8k, 32k), Tencent Hy3 offers a vastly larger capacity of 262k tokens, allowing it to process entire documents without chunking. This can lead to higher accuracy on tasks that require global context. However, smaller models are typically cheaper per token, have lower latency, and may be more efficient for short inputs. For example, a model with a 4k context at a lower price may be better for simple queries. Tencent Hy3’s cost per million tokens is relatively low for its context size, making it a good option for long-context tasks.
Tencent Hy3 is text-only, so it cannot process images, audio, or video. Multimodal models like OpenAI’s GPT-4o can accept such inputs but may have smaller text context windows or higher costs. For text-heavy tasks, Tencent Hy3 may be more cost-effective. If your use case requires understanding visual content, you will need a multimodal model. The comparison depends on whether the added modality justifies the trade-off. Tencent Hy3’s large context window is a distinct advantage for purely textual tasks that involve long documents.
Tencent Hy3 stands out in OrcaRouter's catalog for its combination of a very large context window (262k tokens) and competitive pricing ($0.18/$0.59 per million tokens). It is one of the few models offering such a large context at a sub-$1 output token cost. The model is text-only, which focuses its utility on document analysis and generation. The zero-markup pass-through pricing ensures no surprise fees. This makes it an attractive option for developers and businesses that need to process long texts without breaking the bank, especially compared to other large-context models that may have higher per-token costs.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key=os.environ["ORCAROUTER_API_KEY"],
)
response = client.chat.completions.create(
model="tencent/hy3",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)frequency_penaltyinclude_reasoninglogit_biasmax_tokensmin_ppresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_p| Input / 1M tokens | $0.180 |
|---|---|
| Output / 1M tokens | $0.590 |
| Cache read / 1M | $0.059 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
What developers are saying this week
@misc{orcarouter_hy3,
title = {Hy3 API},
author = {Tencent},
year = {2026},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/tencent/hy3}
}Tencent. (2026). Hy3 API. OrcaRouter. https://www.orcarouter.ai/models/tencent/hy3